Twelve circular structures in Louisiana’s Kisatchie National Forest were reinterpreted as likely World War II military features after field checks found none of the charcoal, tar pipes or clay floors expected at tar kilns.
To detect the rare features, archaeologists trained Mask R-CNN models on simulated kiln-like objects inserted into real lidar terrain data because only 12 known examples existed and their shapes differed from southeastern tar kilns.
The best simulated-data model found all 12 known targets and flagged 11 more for study, but it also produced 686 false positives after filtering; a comparison model trained on real South Carolina kilns found 11 of 12.
The team said the likely military link fits the structures’ location near a former wartime training site, and a training manual suggested they may have been howitzer emplacements.
Researchers said simulated training data can be generated in minutes and may help narrow survey areas faster than pedestrian searches, though automated detection still requires field verification.
When artificial intelligence analyzes terrain, how can archaeologists distinguish between ancient indigenous earthworks and forgotten military training scars?
Will injecting simulated artifacts into landscape data become the ultimate tool for uncovering the world's most elusive archaeological mysteries?